This research was my introduction to neuroscience! In a nutshell, I modelled Perception Biases through Recurrent Neural Networks and analyzed them.
What is Perception bias?
Perception bias or orientation bias in the brain is a well established phenomenon
in Neuroscience where the brain clearly is attracted towards certain orientations
and repelled from certain others. This bias is a specific psychological phenomenon noticed when reporting the orientation of a bar. Humans exhibit a stereotyped pattern of bias during predictions towards ordinal directions as opposed to cardinal ones.
The test is as follows- a screen displays a circular ring, and briefly shows a
red dot at a random angle of on the circumference. The user is required to see this
dot, and recreate the position of the dot once the dot has vanished.
Expected outcome: One would expect the user to make an error in reproducing
the angle/orientation of the red dot 50% of the time above the exact angle and 50%
of the time below the exact angle.
Actual outcome: However, what follows is different. The user’s angle/orientation
predictions are more inclined towards oblique angles (45, 135, 225 degrees) and
away from the cardinal angles (0, 90, 180 degrees).
Goal: To reconstruct the perception bias through recurrent neural networks and
compare the neural activities of this network to another recurrent neural network
that does not have perception bias- and gain insights into evolutionary visual
biases.
Bias vs Orientation Graphs:
Neural Activations during perception bias:
Input vs Output Time Series:
Conclusion:
The neuroscience explanation for the perception bias has been said to
model a ring attractor module, and the goal with the RNN is to produce the
fixed points of the RNN with perception bias. We verify that those points lie along a
ring with higher density towards oblique angles as opposed to cardinal angles.
This project posed questions regarding inherent information biases that exist in the brain, whether the bias is necessary for the optimal working of other functionalities, if this is a by-product of learning, if this could be reconstructed through artificial neural networks while learning altogether another task, etc. To this end, I developed an understanding of how to investigate computational mechanisms by visualizing the topology of fixed points in such networks.
Thanks to postdoc Kaushik Lakshminarasimhan (Center for Theoretical Neuroscience, Zuckerman Institute) for guiding me through this project and exploring it with me!
This analyses led me to continue my neuroscience exploration to another orthogonal study on learning speeds; linked here!
Comentários